def test_greedy(self, graph, path=None):
self.problem.g = to_cuda(graph)
res = [self.problem.calc_S().item()]
pb = self.problem
if path is not None:
path = os.path.abspath(os.path.join(os.getcwd())) + path
vis_g(pb, name=path + str(0), topo='cut')
R = []
Reward = []
for j in range(100):
M = []
actions = pb.get_legal_actions()
for k in range(actions.shape[0]):
_, r = pb.step(actions[k], state=dc(pb.g))
M.append(r)
if max(M) <= 0:
break
if path is not None:
vis_g(pb, name=path + str(j + 1), topo='cut')
posi = [x for x in M if x > 0]
nega = [x for x in M if x <= 0]
# print('posi reward ratio:', len(posi) / len(M))
# print('posi reward avg:', sum(posi) / len(posi))
# print('nega reward avg:', sum(nega) / len(nega))
max_idx = torch.tensor(M).argmax().item()
_, r = pb.step((actions[max_idx, 0].item(), actions[max_idx,
1].item()))
R.append((actions[max_idx, 0].item(), actions[max_idx,
1].item(), r.item()))
Reward.append(r.item())
res.append(res[-1] - r.item())
return QtableKey2state(
state2QtableKey(
pb.g.ndata['label'].argmax(dim=1).cpu().numpy())), R, res
def map_psar2g(psar, hidden_dim=16, rewire_edges=False):
# TODO: if use cluster topology in gnn,
# then rewire_edges=True is required(would be time consuming)
psar.p.reset_label(label=psar.s, calc_S=False, rewire_edges=rewire_edges)
g = to_cuda(psar.p.g)
g.ndata['h'] = torch.zeros((g.number_of_nodes(), hidden_dim)).cuda()
return g
def test_init_state(self, alg, aux, g_i, t, trial_num=10):
init_state = dc(self.episodes[g_i].init_state)
init_states_label = [
QtableKey2state(key) for key in np.random.choice(
self.all_states, trial_num, replace=False)
]
evals = []
for k in range(trial_num):
print('trial:', k)
self.problem.g = dc(init_state)
self.problem.reset_label(label=init_states_label[k])
print('init state:', init_states_label[k])
S = self.problem.calc_S()
bg = dgl.batch([self.problem.g])
state_eval = self.get_state_eval_from_aux(bg, alg, aux)
evals.append(state_eval.detach().item())
print('state_eval:', state_eval)
# print('init S:', S)
for i in range(t):
actions = self.problem.get_legal_actions(
state=self.problem.g, action_type=self.action_type)
S_a_encoding, h1, h2, Q_sa = alg.forward(
to_cuda(self.problem.g),
actions.cuda(),
action_type=self.action_type,
gnn_step=3)
_, r = self.problem.step(state=self.problem.g,
action=actions[torch.argmax(Q_sa)],
action_type=self.action_type)
S -= r.item()
# print(Q_sa.detach().cpu().numpy())
# print('action index:', torch.argmax(Q_sa).detach().cpu().item())
# print('action:', actions[torch.argmax(Q_sa)].detach().cpu().numpy())
# print('reward:', r.item())
# print('kcut S:', S)
print(S)
return init_states_label[torch.tensor(evals).argmax()]
def cmpt_optimal(self, graph, path=None):
self.problem.g = to_cuda(graph)
res = [self.problem.calc_S().item()]
pb = self.problem
S = []
for j in range(280):
pb.reset_label(QtableKey2state(self.all_states[j]))
S.append(pb.calc_S())
s1 = torch.tensor(S).argmin()
res.append(S[s1].item())
if path is not None:
path = os.path.abspath(os.path.join(os.getcwd())) + path
pb.reset_label(QtableKey2state(self.all_states[s1]))
vis_g(pb, name=path, topo='cut')
return QtableKey2state(self.all_states[s1]), res
def test_dqn(self, alg, g_i, t, init_label=None, path=None):
init_state = dc(self.episodes[g_i].init_state)
label_history = init_state.ndata['label'][
self.episodes[g_i].label_perm]
self.problem.g = dc(init_state)
if init_label is None:
self.problem.reset_label(label=label_history[0].argmax(dim=1))
else:
self.problem.reset_label(label=init_label)
S = self.problem.calc_S()
print('init S:', S)
if path is not None:
path = os.path.abspath(os.path.join(os.getcwd())) + path
vis_g(self.problem, name=path + str(0), topo='cut')
for i in range(t):
actions = self.problem.get_legal_actions(
state=self.problem.g, action_type=self.action_type)
S_a_encoding, h1, h2, Q_sa = alg.forward(
to_cuda(self.problem.g),
actions.cuda(),
action_type=self.action_type,
gnn_step=3)
_, r = self.problem.step(state=self.problem.g,
action=actions[torch.argmax(Q_sa)],
action_type=self.action_type)
print(Q_sa.detach().cpu().numpy())
print('action index:', torch.argmax(Q_sa).detach().cpu().item())
print('action:',
actions[torch.argmax(Q_sa)].detach().cpu().numpy())
print('reward:', r.item())
S -= r.item()
print('kcut S:', S)
if path is not None:
vis_g(self.problem, name=path + str(i + 1), topo='cut')
def run_dqn(alg):
t = 0
writer = SummaryWriter(log_folder + save_folder)
for n_iter in tqdm(range(n_epoch)):
T1 = time.time()
if n_iter > len(eps) - 1:
alg.eps = eps[-1]
else:
alg.eps = eps[n_iter]
# if n_iter < 2500:
# alg.gamma = 0
# elif n_iter < 5000:
# alg.gamma = 0.3
# elif n_iter < 7500:
# alg.gamma = 0.5
# elif n_iter < 10000:
# alg.gamma = 0.7
# elif n_iter < 15000:
# alg.gamma = 0.9
T11 = time.time()
# TODO memory usage :: episode_len * num_episodes * hidden_dim
log = alg.train_dqn(target_bg=target_bg
, epoch=n_iter
, batch_size=batch_size
, num_episodes=n_episode
, episode_len=episode_len
, gnn_step=gnn_step
, q_step=q_step
, rollout_step=rollout_step
, ddqn=ddqn)
if n_iter % target_update_step == target_update_step - 1:
alg.update_target_net()
T22 = time.time()
print('Epoch: {}. R: {}. Q error: {}. H: {}. T: {}'
.format(n_iter
, np.round(log.get_current('tot_return'), 2)
, np.round(log.get_current('Q_error'), 3)
, np.round(log.get_current('entropy'), 3)
, np.round(T22 - T11, 3)))
T2 = time.time()
if n_iter % save_ckpt_step == save_ckpt_step - 1:
with open(path + 'dqn_'+str(model_version + n_iter + 1), 'wb') as model_file:
pickle.dump(alg.model, model_file)
t += 1
# with open(path + 'buffer_' + str(model_version + n_iter + 1), 'wb') as model_file:
# pickle.dump([alg.cascade_replay_buffer_weight, [[problem.calc_S(g=elem.s0) for elem in alg.cascade_replay_buffer[i]] for i in range(len(alg.cascade_replay_buffer))]]
# , model_file)
# validation
# test summary
if n_iter % 100 == 0:
test = test_summary(alg=alg, problem=test_problem1, action_type=action_type, q_net=readout, forbid_revisit=0)
test.run_test(problem=to_cuda(bg_hard), trial_num=1, batch_size=100, gnn_step=gnn_step,
episode_len=episode_len, explore_prob=0.0, Temperature=1e-8)
epi_r0 = test.show_result()
if run_validation_33:
best_hit0 = test.compare_opt(validation_problem0)
test.problem = test_problem0
test.run_test(problem=to_cuda(bg_easy), trial_num=1, batch_size=test_episode, gnn_step=gnn_step,
episode_len=episode_len, explore_prob=0.0, Temperature=1e-8)
epi_r1 = test.show_result()
if run_validation_33:
best_hit1 = test.compare_opt(validation_problem1)
test.problem = test_problem2
test.run_test(problem=to_cuda(bg_subopt), trial_num=1, batch_size=100, gnn_step=gnn_step,
episode_len=episode_len, explore_prob=0.0, Temperature=1e-8)
epi_r2 = test.show_result()
if run_validation_33:
best_hit2 = test.compare_opt(validation_problem0)
writer.add_scalar('Reward/Training Episode Reward', log.get_current('tot_return') / n_episode, n_iter)
writer.add_scalar('Loss/Q-Loss', log.get_current('Q_error'), n_iter)
writer.add_scalar('Reward/Validation Episode Reward - hard', epi_r0, n_iter)
writer.add_scalar('Reward/Validation Episode Reward - easy', epi_r1, n_iter)
writer.add_scalar('Reward/Validation Episode Reward - subopt', epi_r2, n_iter)
if run_validation_33:
writer.add_scalar('Reward/Validation Opt. hit percent - hard', best_hit0, n_iter)
writer.add_scalar('Reward/Validation Opt. hit percent - easy', best_hit1, n_iter)
writer.add_scalar('Reward/Validation Opt. hit percent - subopt', best_hit2, n_iter)
writer.add_scalar('Time/Running Time per Epoch', T2 - T1, n_iter)
else:
mode = 'w'
with open(path + 'params', mode) as params_file:
params_file.write(time.strftime("%Y--%m--%d %H:%M:%S", time.localtime()))
params_file.write('\n------------------------------------\n')
params_file.write(json.dumps(args))
params_file.write('\n------------------------------------\n')
# load validation set
if run_validation_33:
with open('/p/reinforcement/data/gnn_rl/model/test_data/3by3/0', 'rb') as valid_file:
validation_problem0 = pickle.load(valid_file)
with open('/p/reinforcement/data/gnn_rl/model/test_data/3by3/1', 'rb') as valid_file:
validation_problem1 = pickle.load(valid_file)
bg_hard = to_cuda(dgl.batch([p[0].g for p in validation_problem0[:test_episode]]))
bg_easy = to_cuda(dgl.batch([p[0].g for p in validation_problem1[:test_episode]]))
bg_subopt = []
for i in range(test_episode):
gi = to_cuda(validation_problem0[:test_episode][i][0].g)
problem.reset_label(g=gi, label=validation_problem0[:test_episode][i][2])
bg_subopt.append(gi)
bg_subopt = dgl.batch(bg_subopt)
for bg_ in [bg_hard, bg_easy, bg_subopt]:
if ajr == 8:
bg_.edata['e_type'][:, 0] = torch.ones(N * ajr * bg_.batch_size)
_, _, square_dist_matrix = dgl.transform.knn_graph(bg_.ndata['x'].view(test_episode, N, -1), ajr+1, extend_info=True)
square_dist_matrix = F.relu(square_dist_matrix, inplace=True) # numerical error could result in NaN in sqrt. value
bg_.ndata['adj'] = torch.sqrt(square_dist_matrix).view(bg_.number_of_nodes(), -1)
greedy_move_history = []
all_states = list(
set([
state2QtableKey(x)
for x in list(itertools.permutations([0, 0, 0, 1, 1, 1, 2, 2, 2], 9))
]))
select_problem = []
for i in range(20000):
# # brute force
problem.reset()
res[0, i] = problem.calc_S().item()
pb1 = dc(problem)
pb1.g = to_cuda(pb1.g)
S = []
for j in range(280):
pb1.reset_label(QtableKey2state(all_states[j]))
S.append(pb1.calc_S())
s1 = torch.tensor(S).argmin()
pb1.reset_label(QtableKey2state(all_states[s1]))
res[1, i] = S[s1]
# print(s1)
# greedy algorithm
# problem.g = gg[i]
pb2 = dc(problem)
pb2.g = to_cuda(pb2.g)
# pb2 = problem
problem.reset()
init_S.append(problem.calc_S())
_, _, sq_dist_matrix = dgl.transform.knn_graph(problem.g.ndata['x'],
1,
extend_info=True)
mat_5by6 = (2 - torch.sqrt(F.relu(
sq_dist_matrix, inplace=True))[0]).numpy().astype('float64')
m_path = os.path.abspath(os.path.join(
os.getcwd())) + '/toy_models/ga_helpers/corr_mat/dqn_5by6.mat'
dump_matrix(mat_5by6, m_path)
# path_m = os.path.abspath(os.path.join(os.getcwd())) + '/ga_helpers/corr_mat/dqn_5by6.mat'
# path_c = os.path.abspath(os.path.join(os.getcwd())) + '/ga_helpers/configs/default.json'
best_solution, acc, best_fitness = run_ga(path_m, path_c)
best_label = np.zeros(k * m).astype('int')
for i in range(k):
best_label[best_solution[i, :]] += i
problem.g = to_cuda(problem.g)
problem.reset_label(label=best_label)
# print('Final S1:', problem.calc_S())
# print('Final S2:', k * m * (m - 1) - k * m * (m - 1) / 2 * best_fitness)
best_S.append(problem.calc_S())
# path = os.path.abspath(os.path.join(os.getcwd())) + '/toy_models/figs/test1'
# vis_g(problem, name=path, topo='cut')
print('init S:', sum(init_S) / num_trial)
print('best S:', sum(best_S) / num_trial)
print('gain ratio:', (sum(init_S) - sum(best_S)) / sum(init_S))
# bg_subopt = dgl.batch(bg_subopt)
# bg_opt = dgl.batch(bg_opt)
#
# if ajr == 8:
# bg_hard.edata['e_type'][:, 0] = torch.ones(k * m * ajr * bg_hard.batch_size)
# bg_easy.edata['e_type'][:, 0] = torch.ones(k * m * ajr * bg_easy.batch_size)
# bg_subopt.edata['e_type'][:, 0] = torch.ones(k * m * ajr * bg_subopt.batch_size)
# bg_opt.edata['e_type'][:, 0] = torch.ones(k * m * ajr * bg_opt.batch_size)
#
# random validation set
# why not generalise to larger graphs?
problem = KCut_DGL(k=k, m=3, adjacent_reserve=8, hidden_dim=h, mode=mode, sample_episode=sample_episode, graph_style='cluster')
test = test_summary(alg=alg, problem=problem, q_net=q_net, forbid_revisit=0)
bg = to_cuda(problem.gen_batch_graph(batch_size=batch_size, style='plain'))
test.run_test(problem=bg, trial_num=trial_num, batch_size=batch_size, gnn_step=gnn_step, episode_len=episode_len, explore_prob=explore_prob, Temperature=Temperature)
test.show_result()
# easy validation set
for beta in [0.1]:
print('beta', beta)
test.run_test(problem=to_cuda(bg_easy), init_trial=1, trial_num=1, batch_size=100, gnn_step=gnn_step,
episode_len=episode_len, explore_prob=0.0, Temperature=1e-8
, aux_model=None
, beta=0)
epi_r1 = test.show_result()
best_hit1 = test.compare_opt(validation_problem1)
j = 0
for i in range(100):
def run_test(self,
problem=None,
init_trial=10,
trial_num=1,
batch_size=100,
gnn_step=3,
episode_len=50,
explore_prob=0.1,
Temperature=1.0,
aux_model=None,
beta=0):
self.episode_len = episode_len
self.num_actions = self.problem.get_legal_actions(
action_type=self.action_type).shape[0]
if isinstance(self.problem.g, BatchedDGLGraph):
self.num_actions //= self.problem.g.batch_size
# if self.action_type == 'swap':
# self.num_actions = self.problem.N * (self.problem.N - self.problem.m) // 2 + 1
# else:
# self.num_actions = self.problem.N * (self.problem.k - 1) + 1
self.trial_num = trial_num
self.batch_size = batch_size
if problem is None:
batch_size *= trial_num
bg = self.problem.gen_batch_graph(batch_size=self.batch_size)
self.bg = to_cuda(
self.problem.gen_batch_graph(x=bg.ndata['x'].repeat(
trial_num, 1),
batch_size=batch_size))
gl = dgl.unbatch(self.bg)
test_problem = self.problem
else:
if aux_model is not None and init_trial > 1:
gg = []
gl = dgl.unbatch(problem)
for i in range(batch_size):
init_label_i = self.test_init_state(alg=self.alg,
aux=aux_model,
g_i=i,
t=episode_len,
trial_num=init_trial)
self.problem.g = gl[i]
self.problem.reset_label(label=init_label_i)
gg.append(dc(self.problem.g))
self.bg = dgl.batch(gg)
self.problem.g = self.bg
test_problem = self.problem
else:
# for validation problem set
self.bg = problem
gl = dgl.unbatch(self.bg)
self.problem.g = self.bg
test_problem = self.problem
self.action_mask = torch.tensor(
range(0, self.num_actions * batch_size, self.num_actions)).cuda()
self.S = [
test_problem.calc_S(g=gl[i]).cpu() for i in range(batch_size)
]
ep = [
EpisodeHistory(gl[i],
max_episode_len=episode_len,
action_type='swap') for i in range(batch_size)
]
self.end_of_episode = {}.fromkeys(range(batch_size), episode_len - 1)
loop_start_position = [0] * self.batch_size
loop_remain_index = set(range(self.batch_size))
self.valid_states = [[ep[i].init_state]
for i in range(self.batch_size)]
for i in tqdm(range(episode_len)):
batch_legal_actions = test_problem.get_legal_actions(
state=self.bg, action_type=self.action_type).cuda()
# if self.forbid_revisit and i > 0:
#
# previous_actions = torch.tensor([ep[k].action_seq[i-1][0] * (self.n ** 2) + ep[k].action_seq[i-1][1] for k in range(batch_size)])
# bla = (self.n ** 2) * batch_legal_actions.view(batch_size, -1, 2)[:, :, 0] + batch_legal_actions.view(batch_size, -1, 2)[:, :, 1]
# forbid_action_mask = (bla.t() == previous_actions.cuda())
# batch_legal_actions = batch_legal_actions * (1 - forbid_action_mask.int()).t().flatten().unsqueeze(1)
forbid_action_mask = torch.zeros(batch_legal_actions.shape[0],
1).cuda()
if self.forbid_revisit:
for k in range(batch_size):
cur_state = ep[k].enc_state_seq[
-1] # current state for the k-th episode
forbid_states = set(
ep[k].enc_state_seq[-1 - self.forbid_revisit:-1])
candicate_actions = batch_legal_actions[
k * self.num_actions:(k + 1) * self.num_actions, :]
for j in range(self.num_actions):
cur_state_l = QtableKey2state(cur_state)
cur_state_l[candicate_actions[j][0]] ^= cur_state_l[
candicate_actions[j][1]]
cur_state_l[candicate_actions[j][1]] ^= cur_state_l[
candicate_actions[j][0]]
cur_state_l[candicate_actions[j][0]] ^= cur_state_l[
candicate_actions[j][1]]
if state2QtableKey(cur_state_l) in forbid_states:
forbid_action_mask[j + k * self.num_actions] += 1
batch_legal_actions = batch_legal_actions * (
1 - forbid_action_mask.int()).t().flatten().unsqueeze(1)
if self.q_net == 'mlp':
# bg1 = to_cuda(self.bg)
# bg1.ndata['label'] = bg1.ndata['label'][:,[0,2,1]]
# bg2 = to_cuda(self.bg)
# bg2.ndata['label'] = bg2.ndata['label'][:,[1,0,2]]
# bg3 = to_cuda(self.bg)
# bg3.ndata['label'] = bg3.ndata['label'][:,[1,2,0]]
# bg4 = to_cuda(self.bg)
# bg4.ndata['label'] = bg4.ndata['label'][:,[2,0,1]]
# bg5 = to_cuda(self.bg)
# bg5.ndata['label'] = bg5.ndata['label'][:,[2,1,0]]
# S_a_encoding, h1, h2, Q_sa = self.alg.forward(dgl.batch([self.bg, bg1, bg2, bg3, bg4, bg5]), torch.cat([batch_legal_actions]*6),
# action_type=self.action_type, gnn_step=gnn_step)
S_a_encoding, h1, h2, Q_sa = self.alg.forward(
self.bg,
batch_legal_actions,
action_type=self.action_type,
gnn_step=gnn_step)
# Q_sa = Q_sa.view(6, -1, self.num_actions).sum(dim=0)
else:
S_a_encoding, h1, h2, Q_sa = self.alg.forward_MHA(
self.bg,
batch_legal_actions,
action_type=self.action_type,
gnn_step=gnn_step)
if beta > 0:
self.problem.sample_episode *= self.num_actions
self.problem.gen_step_batch_mask()
new_bg, _ = self.problem.step_batch(
states=dgl.batch([self.bg] * self.num_actions),
action=batch_legal_actions.view(batch_size,
self.num_actions,
2).transpose(0, 1).reshape(
-1, 2),
action_type=self.action_type,
return_sub_reward=False)
self.problem.sample_episode = self.problem.sample_episode // self.num_actions
self.problem.gen_step_batch_mask()
state_eval = self.get_state_eval_from_aux(
new_bg, self.alg, aux_model)
state_eval = state_eval.view(-1, batch_size).t().reshape(
batch_size * self.num_actions)
self.state_eval.append(
state_eval.view(batch_size, self.num_actions))
Q_sa -= state_eval * beta
terminate_episode = (Q_sa.view(-1, self.num_actions).max(
dim=1).values < 0.).nonzero().flatten().cpu().numpy()
for idx in terminate_episode:
if self.end_of_episode[idx] == episode_len - 1:
self.end_of_episode[idx] = i
best_actions = torch.multinomial(
F.softmax(Q_sa.view(-1, self.num_actions) / Temperature),
1).view(-1)
chose_actions = torch.tensor([
x if torch.rand(1) > explore_prob else torch.randint(
high=self.num_actions, size=(1, )).squeeze()
for x in best_actions
]).cuda()
chose_actions += self.action_mask
actions = batch_legal_actions[chose_actions]
new_states, rewards = self.problem.step_batch(
states=self.bg, action=actions, action_type=self.action_type)
R = [reward.item() for reward in rewards]
# enc_states = [state2QtableKey(self.bg.ndata['label'][k * self.n: (k + 1) * self.n].argmax(dim=1).cpu().numpy()) for k in range(batch_size)]
[
ep[k].write(
action=actions[k, :],
action_idx=best_actions[k] - self.action_mask[k],
reward=R[k],
q_val=Q_sa.view(-1, self.num_actions)[k, :],
actions=batch_legal_actions.view(-1, self.num_actions,
2)[k, :, :],
state_enc=None # enc_states[k]
# , sub_reward=sub_rewards[:, k].cpu().numpy()
,
sub_reward=None,
loop_start_position=loop_start_position[k])
for k in range(batch_size)
]
# # write valid sample episode for training aux model
# tmp = set()
# for k in loop_remain_index:
# if enc_states[k] in ep[k].enc_state_seq[-4: -1]:
# loop_start_position[k] = i
# tmp.add(k)
# else:
# label = ep[k].init_state.ndata['label'][ep[k].label_perm[-1]]
# l = label.argmax(dim=1)
# self.problem.g = dc(ep[k].init_state)
# self.problem.reset_label(label=l)
# self.valid_states[k].append(dc(self.problem.g))
# loop_remain_index = loop_remain_index.difference(tmp)
self.episodes = ep